acceptodds
Under review as a conference paper at ICLR 2027

When Equilibria Bifurcate: Statistical Resolution Limits for Learning Equilibrium Correspondences

Abstract

Learning in games often targets one profile with small exploitability. We show that this can be statistically much easier than learning the complete equilibrium correspondence: in a contiguous family of two-player zero-sum games, a data-independent profile has residual , while every adaptive -query procedure incurs order-one minimax Hausdorff error for the exact correspondence. We therefore interpret approximation tolerance as statistical branch resolution. At a specified base game , we define a branch-survival cost and prove that, at every positive first-order resolution, approximate-equilibrium correspondences converge uniformly to its sublevel sets. The cost equals the duality gap of the perturbation game restricted to the base optimal-strategy polytopes and the support function of a critical-contrast polytope affinely equivalent to the complete base correspondence. Under local alternatives , known-variance Gaussian payoff observations, and deterministic full-support sampling allocations, we obtain a finite-sample risk decomposition with an asymptotically vanishing strategic discrepancy and a local Gaussian random-set minimax limit. This yields a subcritical–critical–supercritical resolution transition and an adaptive lower bound for honest exact-set inference. The statistical comparison is fixed-dimensional and local oracle; complete Hausdorff convergence is proved at positive first-order resolution, while at exact resolution only an outer limit is universal.

Primary area:
learning theory
Keywords:
equilibrium correspondence learning, zero-sum games, multi-agent learning, set-valued statistical inference, local asymptotic minimax theory, variational analysis, uncertainty quantification, Hausdorff loss

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.25 over 4 reviews, above 2% of the others.
Reviewer_4kC9OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

The paper develops a theory for learning the full equilibrium correspondence in finite zero-sum games near bifurcation points. It shows a sharp separation between learning one equilibrium and recovering the entire equilibrium set, and characterizes the statistical resolution limit through perturbation geometry and local minimax analysis.

Critical strengths

The paper studies a gap between finding one equilibrium and recovering the full equilibrium correspondence. It shows that the second task can be much harder near bifurcation points. The paper also develops a coherent theory around this phenomenon, linking perturbation geometry, branch survival, set convergence, and statistical minimax limits.

Critical weaknesses

I do not find the paper well written, and I have several concerns on different aspects:

  1. The paper is technically comprehensive, but the presentation is often overly dense. Many concepts, definitions, and theorem-level claims are introduced in a small space. For example, Section 4 introduces the finite and limiting targets, two minimax risks, strategic and experiment discrepancies, and the Gaussian comparison theorem in rapid succession, leaving relatively little intuition between the definitions and the main result.
  2. The sharper minimax results rely on a local, fixed-dimensional setting with known Gaussian noise and deterministic full-support sampling. These assumptions limit the generality of the statistical conclusions.
  3. The paper is entirely theoretical. Even a small numerical example illustrating the bifurcation and the different statistical resolution regimes would make the results easier to interpret.
  4. I found at least one concrete reference mismatch: the DOI listed for K. Song (2014) (i.e., 10.1016/j.jmva.2013.12.003) appears to correspond to a different article (F. Camerlenghi et al., (2013)). This is concerning because the paper states that the references were checked against primary or publisher sources.
  5. The writing is occasionally not sufficiently rigorous. For example, Table 1 makes theorem-level comparisons using only author-name shorthand, such as “Mills; Altman et al.” rather than explicit citations.

Given the extensive role of AI in literature discovery and manuscript preparation disclosed in the AI Use Statement, I feel that the current version has not yet been sufficiently polished by the authors. Although I appreciate the theoretical completeness of the paper, I believe substantial further effort is needed to improve the manuscript.

Questions for authors
  1. The statistical minimax theory assumes that the base game A⋆A_\star is known. How difficult would it be to construct a globally implementable procedure that first localizes the game and then achieves the same positive-resolution rate?
  2. The exact-set lower bound allows arbitrary adaptive querying, while the matching positive-resolution upper theory is developed for deterministic full-support designs. What is the main obstacle to proving a matching adaptive upper bound?
  3. The Gaussian minimax result is characterized through a random-set decision problem, but the paper does not provide an explicit (near-minimax) decision rule. Can we construct a computationally tractable (minimax) estimator for moderate-size matrix games?
Reviewer_fA8cOpenReview ↗
Rating
1
Confidence
1
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The authors study learning algorithms for game-theoretic problems that do not necessarily find equilibria but approximate equilibria in the sense that the players can not substantially improve their payoffs by deviating from them. While I do know some game theory basics, I have not done research in the field and am not aware of many of the more advanced technicalities, so I cannot summarize the paper contents to a greater extent.

Critical strengths

I am not sure this paper has any discernible strengths.

Critical weaknesses

W1) Starting with the introduction and all throughout the paper, the text is extremely jargon-heavy and does not make any attempt at explaining any of the concepts involved to a general technical audience. Central technical terminology goes undefined and that is pretty unusual even for texts targeted at an expert audience since definitions often don't match exactly across the literature. Given my limited expertise, it is not necessarily alarming that I cannot understand all of the finer technical points, but that it is virtually impossible for me to understand the contribution even on a high level indicates to me that no care has been taken to make this paper fit for human reading.

W2) From reading the paper, it seems that may be (at best) lightly edited LLM output, which also matches the authors' disclosures in the AI statement. There seems to be no effort to make the paper fit for human consumption. It is unclear what value to readers publishing such a paper at a conference would have.

W3) The length of the paper including appendix exceeds 80 pages, and almost all of it is dense mathematical/technical material which by the authors' own disclosure is all AI generated. There is no discussion of why such an unusual amount of technical material is necessary to establish the technical results, and what value publishing it would bring to the theoretical understanding of online learning. After all, the publications at ICLR are still intended mainly for human reading. If the value is just establishing the technical validity itself with no insight into methodology or technique (at least for humans), it stands to reason that the proofs should also be published in a machine-friendly form (e.g., a theorem checker language). In any case, the 2.5 weeks review process of ICLR is in my view not remotely suitable to determine if a paper of this length should be published. In any case, there is no way here for me to say anything about technical soundness.

Questions for authors

Q1) What are the 2-3 most significant contributions of the paper? I did not find a contribution list. Q2) What are examples of practical problems that this theoretical contribution might have potential impact on in the future, respectively, the theoretical understanding of which it could potentially advance? Q3) Is it possible to run numerical experiments on such problems? If not, discuss the limitations and sketch the path of research that will be needed to close that gap.

Reviewer_fArbOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The article claims to study learning in games. Unfortunately, it is difficult to give any farther summary of the submission, due to extremely technical writing. Please refer to the critical weaknesses of the article for more details.

Critical strengths

None

Critical weaknesses

The article is close to impossible to read. It has been clearly fully written by AI without any effort from the authors to make it understandable. Unfortunately, submissions like this one cannot be recommended for acceptances in a conference such as ICLR.

The writing issues start already on the abstract, where terms such as "exploitability", "residual", "branch resolution", "sublevel sets", and others, are never explained. Although this can always be the case in an abstract where the space is limited, the same issues arrive during the introduction. Some examples are

  • Line 34 : "Exploitability"
  • Line 46 : "without data" is mentioned but no context is given
  • Line 48 : "branch completeness" is never defined
  • Line 51 : E0\mathcal{E}_0 is not defined. I suppose it corresponds to the set of Nash equilibria of the game AA, but still.
  • Figure 1 and Table 1 are never referenced on the document nor discussed
  • Line 71 : "correspondence visible at resolution"
  • Line 79 : C∗\mathcal{C}_{*} is not defined
  • Line 82-83 : "equilibrium multiplicity is encoded exactly as the dual geometry governing first-order persistence" is not understandable
  • Line 88 : "epi-derivative" is not explained
  • Line 127 : "global query model" is not defined

Unfortunately, already in Page 3 the reader is unable to understand the problem addressed by the article, the objectives, the contributions, even the formal setting. Due to all this issues, I can only recommend the rejection of the article.

Questions for authors

I have no questions.

Reviewer_D2LDOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

To be honest, even the abstract is extremely difficult to understand, and I could barely understand what the paper is trying to investigate overall. My understanding is that the paper investigates the difference in difficulty between finding a single approximate equilibrium and estimating the entire equilibrium set, but I could not understand the problem setting or the contribution of the paper.

Critical strengths

None

Critical weaknesses

The most significant weakness of this paper is its lack of clarity.

For example, even the abstract fails to convey what the paper investigates. The second sentence introduces "contiguous family", "residual" and "minimax Hausdorff error" in the same sentence, while the third sentence says that interpret approximation tolerance as statistical branch resolution. However, the connections between these concepts and what some of these words mean are unclear. Furthermore, the discussion proceeds without explaining the intuitive meaning of "branch-survival cost" or the definition of HH. The word “subcritical–critical–supercritical resolution transition" is also introduced without specifying what they are. I cannot even understand the abstract.

There are also serious problems with the appendix. Appendices A.3 to A.5 consist largely of lists of mathematical facts, and their connections to the main body of the paper are unclear. More seriously, Appendix L.2 contains sentences of the authors' manuscript preparation and verification process, including, for example, a description of a directory named "verification", for which no supplemental material has been provided.

These issues raise serious doubts about whether the manuscript has been adequately reviewed, and I do not believe that a manuscript in its current state is suitable for publication.

Questions for authors

None

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.